Bootstrapping Road Sign Detection for Self-Driving Cars using Weakly-Supervised Learning

Given the escalating trend in the number of cars on the public roads, advanced autonomous driver-assistance systems become more accessible to the masses in order to keep safe traffic participants. This work addresses the problem of traffic signs detection constrained by running in a minimal embedded platform. Our solution consists of the generation of a synthetic object detection dataset using CARLA Simulator, a popular self-driving car virtual environment, enhancing it with image augmentation policies and bootstrapping the model performance by using knowledge distillation from a model ensemble. We make use of modern weakly supervised techniques to minimize labelling noise and achieve a fast, predictive, high-precision model that performs well in real-life scenarios, having a mean average precision of over 53%. The model is integrated with ease into real-time applications achieving 19 FPS on our embedded platform that uses a small size Coral Edge TPU USB Accelerator. Our proposed computer vision solution that powered a scaled self-driving car enabled the team to rank third in the Bosch Future Mobility Challenge 2022, an IEEE ITSS certified contest that encourages the development of complete autonomous driving solutions for scaled vehicles in controlled real-life scenarios.